Model Council — models compete on your prompt

The Council pane answers “which model does this best?” with a contest instead of a vibe. Pick contestants, pick a format, submit the prompt once, and watch scored results come back.

Formats

1. Arena — every contestant answers the same prompt in parallel; the judge scores all of them.
2. Tournament — head-to-head brackets; winners advance until one remains.
3. Debate — models argue positions across rounds before judgment.

Judging

The judge is blind — it scores answers without knowing which model produced them — and rates per dimension (correctness, clarity, and friends), so the radar chart shows where a model wins, not just that it won. Code and HTML answers get an artifact preview so you can eyeball the output itself.

Steps

1. Open it with open model council.
2. Pick contestants and a judge model.
3. Choose Arena, Tournament, or Debate; enter the prompt; run.
4. Read the radar + per-dimension scores; open artifacts to inspect answers.
5. Save the setup if you’ll rerun it — saved councils reload contestants, judge, and format in one click.

Good to know

• CLI contestants run your real local agents, so their answers reflect exactly what you’d get day to day.
• For repeatable scored suites (rather than one-off prompts), use the Benchmark suite.